sgl-project/sglang · error · ValueError

Unsupported Kimi-K3 encoder media item: {image}

Error message

Unsupported Kimi-K3 encoder media item: {image}

What it means

prepare_kimi_k3_encoder_inputs accepts per-image items either as raw images or as dicts of the form {'type': 'image', 'image': <payload>, ...}. Any dict that does not have type == 'image' and an 'image' key is rejected with this ValueError, protecting the encoder from unsupported media kinds (e.g. video/audio dicts or malformed entries).

Source

Thrown at python/sglang/srt/multimodal/kimi_k3_image_processing.py:83

        "in_patch_limit",
        "patch_limit_on_one_side",
        "image_mean",
        "image_std",
    )
    missing = [name for name in required if name not in media_proc_cfg]
    if missing:
        raise ValueError(
            "Kimi-K3 image processor is missing deferred-preprocessing config: "
            + ", ".join(missing)
        )

    concrete_images = []
    content_digests = []
    for image in images:
        content_digest = None
        if isinstance(image, dict):
            if image.get("type") != "image" or "image" not in image:
                raise ValueError(f"Unsupported Kimi-K3 encoder media item: {image}")
            content_digest = image.get("content_hash")
            image = image["image"]
        concrete_images.append(image)
        content_digests.append(content_digest)

    patch_size = int(media_proc_cfg["patch_size"])
    merge_kernel_size = int(media_proc_cfg["merge_kernel_size"])
    deferred_preprocessing = functools.partial(
        KimiK3DeferredPreprocessing,
        backend="gpu" if use_gpu_preprocessing else "cpu",
        image_mean=list(media_proc_cfg["image_mean"]),
        image_std=list(media_proc_cfg["image_std"]),
        transparent_bg_config=media_proc_cfg.get("transparent_bg_config"),
    )

    items = []
    grids = []
    original_image_sizes = []

View on GitHub (pinned to 0132848349)

Solutions

1. Filter the content list to items with item.get('type') == 'image' before passing
2. Ensure each dict has an 'image' key holding the image payload; unwrap raw images if dicts are unnecessary
3. Normalize other modalities through their own encoder-input preparers instead of this image API

Example fix

// before
images = content  # contains {'type':'text',...}
prepare_kimi_k3_encoder_inputs(images, cfg)
// after
images = [c for c in content if c.get('type') == 'image']
prepare_kimi_k3_encoder_inputs(images, cfg)

def is_kimi_k3_image_item(x):
    return not isinstance(x, dict) or (x.get("type") == "image" and "image" in x)

try:
    prepare_kimi_k3_encoder_inputs(images, cfg)
except ValueError as e:
    if "Unsupported Kimi-K3 encoder media item" in str(e):
        images = [i for i in images if is_kimi_k3_image_item(i)]
        prepare_kimi_k3_encoder_inputs(images, cfg)
    else:
        raise

When it happens

Trigger: Passing an images list containing a dict like {'type': 'video', ...}, {'type': 'image'} without an 'image' key, or a dict with a differently-named payload key (e.g. 'url' instead of 'image').

Common situations: Forwarding an unfiltered chat content list (which contains text/video entries) directly as images; renaming keys when building the request payload; partial refactors of the mm item schema.

Related errors

• mm_content_hashes has {len(content_hashes)} entries for {med
• Unsupported image type: {type(image)}
• When using multiple prompts with multiple input images, prov
• {key}.position_ids is required
• {key}.{field} is required

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5931e8a3208189aa. Report an issue: GitHub.